Request a review

AI news6 min read

OpenAI releases GPT-6.2 Astra and Sol at lower cost per task

OpenAI’s new frontier pair arrives in ChatGPT Work, Codex and the API. Astra gets 20% cheaper; Sol nearly matches it on coding at a fifth of the cost.

In this story9 parts

OpenAI has released GPT-6.2 Astra and GPT-6.2 Sol, the next versions of its two frontier models. OpenAI describes Astra as its model for the hardest projects and Sol as an efficient model for everyday work and multistep agent tasks. The theme of the launch is cost: OpenAI says Astra does more for less than a third of GPT-6 Astra’s cost per task, and Astra’s API price falls 20%.

It comes less than two months after OpenAI cancelled the GPT-6.1 Astra release over safety concerns, so GPT-6.2 Astra is the first Astra upgrade since GPT-6 Astra in September.

OpenAI’s launch video

OpenAI’s announcement on X: “GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world.”

Where you can use it

  • ChatGPT Work and Codex: for Plus, Pro, Enterprise and Edu users. OpenAI staff say the rollout does not add either model to regular ChatGPT chat, and availability depends on your plan, workspace settings and rollout progress. Update the desktop app and the Codex CLI if you do not see them.
  • Ultrafast: Pro 500 subscribers can use Ultrafast mode with both models. It is coming to Pro 200, starting with Sol.
  • API: gpt-6.2-astra and gpt-6.2-sol, on the Responses, Chat Completions and Batch endpoints.

Price

From OpenAI’s model pages, per million tokens:

GPT-6.2 Astra GPT-6.2 Sol
Input $8 $2
Cached input $0.40 $0.10
Output $40 $10
Context window 1.05M tokens 1.05M tokens
Max output 128K tokens 128K tokens
Knowledge cutoff May 18, 2026 June 1, 2026

Batch requests cost half. Fast mode costs twice the standard rate, and Ultrafast three times until February 28, 2027, then six times from March 1. For Sol, prompts over 272K input tokens cost double for input and 1.5 times for output. Astra is 20% cheaper than GPT-6 Astra; Sol’s price is unchanged from GPT-6.1 Sol. Both take text and images in and return text.

The benchmarks: what OpenAI claims

OpenAI’s announcement presents most results as gains over earlier models, with the full scores in charts:

  • Agents’ Last Exam: Astra at xhigh effort scores 5.2 points above GPT-6 Astra at max effort, at less than half the cost per task. The Decoder reports that OpenAI’s chart puts Astra 5.5 points ahead of Claude Fable 5.2 at a fraction of its cost.
  • FrontierCode 2.0 (coding): Sol comes within 0.5 points of Astra on merge-ready changes to real codebases, at about one-fifth of the cost. Fable 5.2 still posts the highest peak score in OpenAI’s chart, at several times the cost per task.
  • GDPval-v2 (knowledge work): Astra gains 6.2 points on GPT-6 Astra while using 22% fewer output tokens, per The Decoder’s reading of the post.
  • Terminal-Bench-Science: Astra gains 9 points on GPT-6 Astra.
  • ExploitBench Internal Port (cybersecurity): Astra gains 22 points on GPT-6 Astra. OpenAI calls it its most capable cyber model.

These are OpenAI’s own evaluations.

The independent numbers

Outside testers had results on launch day:

Benchmark GPT-6.2 Astra GPT-6.2 Sol Claude Fable 5.2 Source
Artificial Analysis Intelligence Index v5 62 57 65 Artificial Analysis, via The Decoder
AA-Robotics v0.1 (real robot arm) 79 78 75 Artificial Analysis
AA-3D (interactive 3D scenes, Elo) 1,318 1,259 1,271 Artificial Analysis, via The Decoder

On the Intelligence Index, Astra is second behind Anthropic’s Claude Fable 5.2, five points above GPT-6 Astra, with Gemini 4 Carbon (60) and Claude Opus 5.5 (59) behind it. Sol matches GPT-6 Astra’s original score at roughly 15% of its cost per task. Artificial Analysis also says humans preferred Astra’s 3D scenes in 139 of 152 head-to-head votes.

Not everything moved up. The Decoder reports that Artificial Analysis found no gain for Astra on its coding-agent index and a three-point drop for Sol, and measured hallucination rates of 59% (Astra) and 55% (Sol) on its hard factual test.

OpenAI also published a demo of Astra building 3D sets for video:

Safety

OpenAI says the models launch with tighter safeguards for higher-risk cybersecurity help, and it is working to bring them to Daybreak so approved defenders get fewer refusals. According to The Decoder’s reading of OpenAI’s safety material, Astra is the first model OpenAI has released externally that reaches the “High” threshold for AI self-improvement in its Preparedness Framework, and both models reach the “Critical” cybersecurity threshold, as GPT-6 Astra did.

What developers should know

OpenAI’s GPT-6.2 guide flags two changes:

  • Reasoning is dropped on history edits. If you change earlier turns, switch to a smaller model, or change the end user or organization mid-conversation, the API now drops the model’s earlier reasoning, which can lower quality and cost more tokens. Keep the conversation prefix unchanged to keep reasoning and prompt caching.
  • More concise by default. Astra writes more tersely and may stop to propose a next step when a request is unclear. Revisit old “be concise” instructions, and tell it when it may keep working on its own.

Early users in OpenAI’s developer forum report mixed impressions, including stiffer writing and style instructions that do not stick. Sam Altman has reportedly said there is more work to do on making the models “nice to work with.”

What it means for working professionals

  • Engineers: try Sol first. On OpenAI’s own coding test it is within half a point of Astra at a fifth of the price, and independent coding-agent scores did not improve for Astra.
  • Analysts, lawyers and consultants: Astra’s gains are biggest in long agent tasks and knowledge work. Check its facts; independent hallucination rates remain high.
  • Budget holders: the per-token price fell for Astra, but Ultrafast costs three times the standard rate now and six times from March. Measure cost per finished task.

We have not tested GPT-6.2 yet. Everything above comes from OpenAI’s announcement, model pages, guide and videos, and the independent results named in each section. For OpenAI’s agent workspace where the models launch, see our ChatGPT Work review.

Sources

Every story names its source. Check our work.